The Cheese Plant Running at 3% Capacity Is the Real Story of Salesforce AI Agents
Salesforce turned a 97%-idle cheese plant into a sales-and-fulfillment machine. The real question is what that trick requires — and who it leaves out.
Petaluma Creamery's plant was built to process 140,000 pounds of cheese per day. It is currently running at roughly 3 percent of that capacity. When Salesforce and its Agentforce platform landed the story that Fortune ran as a cover, the framing was rescue: AI saved a 113-year-old family cheese business from extinction. That is accurate. It is also a story about a facility that is still, in industrial terms, almost entirely idle — and the question of whether that matters.
The short answer is no, not yet. The longer answer is that the AI agents are not making cheese. They are making the economics of a 97-percent-idle plant worth tending.
What the agents actually replaced
Before Daniel Peter joined as CTO — he spent 17 years in the Salesforce ecosystem before coming aboard — Petaluma ran on paper and tribal knowledge. Orders arrived on handwritten forms. QuickBooks entries used a 150-SKU code hierarchy that employees memorized: yellow cheddar was "C:CY." Every transaction required manual conversion from cases to pounds. There was no fiber internet, just T1 connections that were common in the 1990s.
"Forget about AI," Daniel said. "It was like, how do we have a digital representation of an order?"
The deployment did not begin with agents. It began with data foundation: 20-plus years of email archives, QuickBooks invoices, and lab data for dairy product testing. Daniel's sequence was establish the data layer first, then automate, then add intelligence. The agents came last.
What they do now: order-to-cash replaced the memorized SKU hierarchy with natural language search — type "Firehouse" or "Jack" and the right product surfaces, with case-to-pound conversion handled automatically. Predictive ordering ingests customer history and pre-populates reorders, surfacing items a buyer typically orders but forgot to mention. Delivery routing accepts geographic constraints as plain English prompts rather than code. A routing update means editing a sentence, not spinning up a development cycle. Milk traceability tracks every gallon from farm arrival through production batches to lot numbers on store shelves, with state compliance reporting built in.
The result: active accounts grew from 13 to more than 300. Benu, the three-Michelin-star restaurant in San Francisco, signed up. The Sacramento Kings now stock Petaluma products in their arena concession stands.
Next up is a fully agentic customer service layer that queries the creamery's full 20-year data archive — faster than any human rep, and more exhaustively. "A person is not going to search 20 years of data and come up with a nice customer service answer," Daniel said. "They are going to just stop at the first thing they find."
The prerequisites problem
That is where the story becomes less triumphant and more instructive.
The AI layer only worked because the data existed. Twenty years of clean digital records is the prerequisite the Salesforce case study page does not highlight. Most small food manufacturers do not have it. They have shoeboxes full of receipts and knowledge that walks out the door when a bookkeeper retires. The creamery also had a CTO with 17 years of Salesforce experience and the capital to run fiber and migrate to the cloud before adding an AI layer.
Small food manufacturers without those prerequisites cannot replicate this. The story demonstrates what Agentforce can do for the subset of small manufacturers that already had the foundational infrastructure. It does not demonstrate democratization.
There is also the question of what the agents are actually doing. They are handling order entry, predictive restocking, routing, and traceability. These are real productivity gains. They are not making cheese. The plant is still running at 3 percent of nameplate capacity.
Petaluma peaked at roughly $50 million in annual revenue, supplied hundreds of Chipotle locations, and had its cheese served at the Kentucky Derby. Julia Child called it the best white cheddar she ever ate. After COVID, the death of Larry Peters' father, his own open-heart surgery, and a buyer who strung them along for 18 months promising $50,000 a month in rent that never appeared, they had 13 active accounts.
The Salesforce deployment changed that. The $10 million revenue target for next year is a real number against a real near-zero baseline. The long-term vision of $200 million to $300 million assumes the plant scales toward the 140,000 pounds per day it was built for. Whether that happens depends on whether the AI layer can expand the customer base faster than the artisanal product mix allows, and whether the data foundation holds at volumes the business has not seen in decades.
The Salesforce angle
Salesforce is not running this case study out of altruism. Every enterprise software company is looking for proof that agentic AI works outside Fortune 500 IT departments. A 113-year-old creamery with a founder who once sold cheese out of a woodshed is more compelling than a deployment at a tech-forward retailer.
The broader question — whether AI agents can unlock value from stranded industrial capacity at scale — is real. The gap between Petaluma's 3 percent utilization and its nameplate potential is enormous. If the playbook generalizes, it matters for thousands of similar facilities. If it requires 20 years of clean data and a Salesforce veteran to implement, the generalization stops at companies that already look like Salesforce customers.
The plant is still idle. The agents are not a cheese story. They are a sales-and-fulfillment story wearing cheese as a costume.